forge

{
  "name": "forge",
  "version": "0.9.0",
  "description": "Agent-engineering workflow skills for Claude Code, Codex, and opencode — twenty-two fg-* skills: the four-stage loop (fg-ask → fg-run → fg-learn → fg-done) plus outside-the-loop utilities, including fg-security (vendored security audit) and fg-debug (vendored diagnosis only: active failures return to fg-run; independent diagnoses route by size, durable eval, and all-exit safe cleanup).",
  "author": {
    "name": "gyuha",
    "email": "[email protected]",
    "url": "https://github.com/gyuha"
  },
  "homepage": "https://github.com/gyuha/forge",
  "repository": "https://github.com/gyuha/forge",
  "license": "MIT",
  "keywords": [
    "forge",
    "workflow",
    "planning",
    "codex",
    "claude-code"
  ],
  "skills": "./skills/",
  "interface": {
    "displayName": "Forge",
    "shortDescription": "Plan, execute, learn, and seal agent work",
    "longDescription": "A shared Claude Code and Codex workflow for turning one task into a reviewed plan, verified execution, retained learning, and a sealed result.",
    "developerName": "gyuha",
    "category": "Productivity",
    "capabilities": [
      "Interactive",
      "Write"
    ],
    "websiteURL": "https://github.com/gyuha/forge",
    "defaultPrompt": [
      "Start a new task with Forge.",
      "Show the current Forge status.",
      "Run the next Forge step."
    ],
    "brandColor": "#D97757"
  }
}

{
  "name": "forge",
  "description": "An agent-engineering workflow skill set (fg-*) for Claude Code. Includes fg-security, a codebase security audit (methodology vendored from cloudflare/security-audit-skill, MIT) whose artefacts stay outside the repo and whose severity-gated findings become fix-forward plans, and fg-debug, a conversational diagnosis discipline for hard bugs (vendored from mattpocock/skills' diagnosing-bugs, MIT) — diagnosis only and state-first: an active failed task returns to fg-run, while only independent diagnoses branch to fg-quick or an approved fix-forward plan; a red command counts as a persistent eval only when durable, raw captures stay outside the repo, every exit cleans temporary artefacts, and unattended drives always skip it (ADR 260907-140655). A development loop that takes one task through a single cycle of ask·plan → execute → retro → done. Consists of fg-ask (inquiry/triage + grill-with-docs-style conversational grilling), fg-run (backlog selection menu · Run all, Dynamic Workflow execution), fg-learn (retro · doc promotion · eval promotion), and fg-done (done · seal · re-run guard — the mechanical seal runs as the deterministic forge-done.sh/.js script shared by every seal path, ADR-0030), plus fg-map — an on-demand utility outside the loop that maps the codebase into .forge/codebase/ with parallel subagents so grilling reads a map instead of re-exploring the code (cuts context rot), and fg-quick — a lightweight lane outside the loop for trivial tasks that still grills lightly but skips formal artifacts (no ADR/plan/retro), records one line to .forge/quick/LOG.md, runs directly, and bails to fg-ask if the task turns out non-trivial, fg-status — a read-only reporter that surveys .forge/ and prints where every task stands plus the single next step you need (writes nothing, never auto-runs), fg-next — derives that same next step and runs it, announcing the step then invoking the skill rather than only reporting it (one-shot, except that a retro concluding without a re-grill recommendation chains straight into the seal — ADR-0026; fg-status reports, fg-next acts), with an 'all' mode (fg-next all) that drives backlog tasks to completion until empty, auto-progressing mechanical steps and halting at conversational walls, and fg-config — the unified settings surface for .forge/config.json (outside the loop), reading and writing all six keys — simple (auto-seal mode: after UAT passes, fg-run unconditionally skips the retro as retro: skipped (simple mode) and chains into the deterministic seal in the same turn; the verification gate ADR-0009 is inviolable, so verified: pending|failed is never auto-sealed), eco (economical delegated execution: fg-run caps its delegated workflow subagents at sonnet — lower only, the session model is untouched — and the embedded Eco laziness-first discipline in skills/fg-config/ECO.md (code simplicity plus terse-communication output, caveman-derived) is injected into fg-run subagents, woven into fg-ask grilling as a YAGNI lens, and adopted by the current session; eco also replaces task-end handoffs — fg-run's handoff, fg-done's explicit single seal, and the batch/unattended paths — with a compact summary table plus a per-slice table, leaving execution-time narration untouched — ADR-0014, ADR 260730-230321), tdd (a test-first default: fg-ask then defaults to asking, fg-run runs test-first — ADR-0008), driveCommit/driveCommitMessage (per-task local commit on unattended drives), and defaultBranch — replacing the former per-key toggle skills and absorbing their triggers (ADR 260905-212045), and fg-loop — a goal-driven momentum loop with bounded replan: an initial conversational inquiry pins a machine-verifiable stop condition, an authorized fix-forward replan scope, and a replan cap (default 3 rounds) into .forge/loop.md, then it drives tasks unattended (run → UAT → auto-skip retro → seal, auto-selecting recommended defaults at every soft decision and asking the user nothing) until the checks pass, generating in-scope fix-forward tasks within the cap and halting only at the walls (unverifiable UAT, genuine fork, cap exhausted, no progress, check-tension oscillation, safety/irreversible action — ADR-0016). On a non-default git branch the whole forge state moves under .forge/branch/<branch>/ (git-tracked, namespaced so parallel branches never collide on .forge state — ADR-0011), and fg-merge — an outside-the-loop utility that, after you git merge a branch (or via fg-merge <branch>, which runs that git merge for you — interactive, default branch, ADR 260717-10a), integrates that branch's .forge/branch/<branch>/ into .forge/ via the deterministic forge-merge.sh/.js script (so it also works AI-free in CI): moving time-based ADR IDs as-is with an exact-ID clash taking the next free letter (no cascade renumber), merging CONTEXT terms, folding done history, removing the branch folder; it halts non-zero on genuine conflicts (a redefined term, an incoming NNNN collision), leaving semantic ADR contradictions to PR review, and fg-cleanup — an outside-the-loop utility that retires stale or superseded ADRs to .forge/adr/retired/ on your approval, keeping the active decision set current without deleting the why (ADR-0012), and fg-statusline — an outside-the-loop setup utility that shows the current loop progress in your terminal statusline in one of two modes: append (method 1) auto-wraps your existing statusLine so forge shows as an extra row rather than replacing it (ADR-0017), or merge (method 2) installs a forge-owned unified script rendering daleseo-style system info (model·effort·dir·git, plus context/usage/weekly usage bars) alongside forge progress in one command (ADR-0029), and fg-adversarial-review — an outside-the-loop utility that runs an optional adversarial review between fg-run and fg-learn: it assumes the result is wrong and fans out six lenses (failure points, hidden assumptions, misread requirements, security/performance/data-loss, unexpected misuse, weak decisions) as parallel workflow subagents, records findings to .forge/review.md, and on your approval turns fix-needed findings into a fix-forward plan to run again — optional, never a seal gate, always auto-skipped in unattended drives (ADR-0018), and fg-doctor — a read-only integrity health check (outside the loop) that surveys the .forge/ state contract and the docs/manifests and reports inconsistencies with a severity and an actionable fix hint each (orphaned run.md, broken/missing STATUS fields, slug-pairing mismatches, half-sealed done/, an orphaned branch root you forgot to fg-merge, manifest version drift across the 3 places, README bilingual drift, skills missing from CLAUDE.md), now script-backed by forge-doctor.sh/.js and usable AI-free as a CI gate; it writes nothing and never auto-fixes — fg-status reports where you are, fg-doctor reports whether the state is healthy (ADR-0019), and fg-help — a read-only usage-help reporter outside the loop: /forge:fg-help prints an overview of every forge fg-* skill grouped by the loop's four stages plus the outside-the-loop utilities, and /forge:fg-help <command> prints a short 4-line usage detail (what it does · when to use · trigger · next step) for one skill; it reads each skill's own SKILL.md description as the single source (no copied usage text, so a new skill shows up automatically) and renders the output in the user's language at runtime, so it is LLM-executed with no deterministic script twin — the translation is why it differs from fg-status/fg-doctor; it writes nothing and never auto-runs (fg-status: where am I, fg-doctor: is the state healthy, fg-help: how do I use these) (ADR 260814-104534), and fg-drop — an outside-the-loop utility that discards incomplete work you no longer want (a backlog plan, the active slot, an awaiting-retro task in executed/, or a halted goal loop): it lists the items with a per-item risk level (a checkbox dialog for ≤4, a numbered text list for ≥5), then a separate question to hard-delete (default, no trace) or archive to .forge/dropped/, behind a confirmation gate that warns already-run code is not reverted; it removes forge state only — never git, never your code — and a halted goal loop is dropped whole with its member tasks excluded from individual drop (ADR-0021), and fg-agents — an outside-the-loop utility that generates a project's domain agents by conversational grilling (outside any workflow): it derives the recurring, separable roles in the project and writes standard Claude Code subagent cards to .claude/agents/<role>.md, each description carrying a 'when to use' so fg-run can dispatch a matching role as agentType; the cards load at session start so they take effect only after a restart, and the whole path is graceful — a project with no domain agents runs exactly as before (ADR-0024), and fg-showme — an outside-the-loop utility that opens a browser-based visual companion in any conversation (vendored from obra/superpowers v6.1.1's Visual Companion with MIT attribution): a zero-dependency local Node server (scripts/server.cjs plus 4 companion files) shows mockups, diagrams, and visual A/B options the agent pushes as HTML in a live browser tab and takes the user's answers back as JSONL events — a click settles a choice and a screen can carry a text box — which the agent reads back; session files live under the top-level .forge/showme/<session>/ on every branch (volatile, gitignored — a global exception like config.json and codebase/, never the branch root), with session-key URL auth, restart/reconnect survival, a 4-hour idle auto-shutdown, and no telemetry or remote assets; during fg-ask grilling the companion is offered once just-in-time at the first visual question (a standalone message, never re-offered after a decline), fg-ask consuming ../fg-showme/VISUAL.md and scripts/ by file reference (the single-definition convention) and shutting the server down at handoff, while this skill is the standalone entry point (fg-showme to start, fg-showme stop to shut down) (ADR 260719-224442), and fg-agenda — an outside-the-loop planning utility that gives forge a home for a decision not yet made: everything else forge produces is a build artifact and an ADR records a decision already made, so work still in fog either gets forced into a prematurely-complete plan or its open decisions scatter across conversation and retro prose and vanish; fg-agenda settles the destination with you, grills breadth-first to surface what must be decided, and keeps one file .forge/agenda.md (destination · decided · open questions · fog · out of scope), resolving one question at a time by fg-ask's grilling method until the way is clear and then deleting itself, the durable trace being the ADRs and backlog plans it produced — the agent finds what must be decided and the human answers every question (it never answers its own), a decision that becomes buildable leaves the agenda to be an ordinary backlog plan through fg-ask so the agenda is never a second backlog, and it stays out of the fg-next/fg-status next-step chain (fg-status reports a present agenda in one line); adapted in concept, not code, from the Wayfinder skill in mattpocock/skills (MIT) (ADR 260805-201313). Beyond the skills, the plugin ships one hook (auto-discovered hooks/hooks.json — installed with the plugin, no settings.json edit): at SessionStart (startup/resume/clear/compact) it injects a short notice when, and only when, an unsealed tail exists — a task that ran but was never sealed, a task parked awaiting retro, or a halted goal loop — so the agent surfaces it before any new work starts; it never auto-runs and never auto-seals (the decision stays the user's), stays completely silent on a clean repo or a backlog-only queue, dispatches through a bash-then-node polyglot wrapper that exits silently when neither runtime exists, and — because hooks load at session start — applies from the NEXT session (ADR 260727-201031). Use in contexts like 'start a new task', 'start with forge', 'let's work on this', 'refine the plan', 'map the codebase', 'forge quick', 'forge status', 'forge next', 'forge loop', 'forge config', '설정', 'simple mode', 'eco on', 'tdd on', 'forge merge', 'forge cleanup', 'forge statusline', 'adversarial review', '적대적 리뷰', 'forge doctor', '무결성 검사', '상태 점검', 'forge drop', '작업 버리기', '계획 지워', 'forge agents', '도메인 에이전트 만들어', 'create project agents', 'forge visual', 'visual companion', '목업 보여줘', 'forge agenda', '의제', '결정 대기열', 'forge help', 'fg-help', '도움말', '사용법', 'debug this', '디버깅해줘', '버그 잡아줘', 'hunt this bug'.",
  "version": "0.9.0",
  "author": {
    "name": "gyuha",
    "email": "[email protected]",
    "url": "https://github.com/gyuha"
  },
  "homepage": "https://github.com/gyuha/forge",
  "repository": "https://github.com/gyuha/forge",
  "license": "MIT",
  "keywords": [
    "forge",
    "workflow",
    "loop",
    "grill-with-docs",
    "planning",
    "retrospective",
    "claude-code"
  ]
}